Fast and Robust Training and Deployment of Deep Reinforcement Learning Based Navigation Policy

摘要

The application of Deep Reinforcement Learning (DRL) as a comprehensive solution to automobile control and navigation has seen extensive exploration. Essentially, the agent refines the control policy, which maps the robot's actions from its inputs, by interacting with the environment. Despite this, the practical implementation of DRL faces significant challenges due to the prolonged training time required and the sim-to-real disparity. In this study, we trained a DRL-based motion planner under carefully designed simulation scenarios capable of emulating physical interaction with humans. To address the mentioned challenges, human experiences were incorporated into the experience buffer domain to expedite the convergence of training. It could also enable the agent to acquire effective strategies for interaction with dynamic objects. Furthermore, to enhance the robustness of the system, we introduced noise into both the input and target action networks. The efficacy and generalizability of our training framework were validated using an Ackermann mobile robot equipped solely with an economical 2D LiDAR and odometry. Both ROS Gazebo simulations and real-world experiments demonstrated the high performance and efficiency of the proposed approach, underscoring its potential utility in real-world navigation scenarios. The video link is: https://www.youtube.com/@TAOWeizhi/playlists.